The $4.1 Trillion Question: Who Governs the Agentic Economy?

The Economy That Runs Itself

In March 2026, a mid-tier European logistics company discovered something unusual during a routine audit. Over the preceding six weeks, its AI procurement agent had renegotiated seventeen supplier contracts, switched three freight providers, hedged currency exposure across four corridictions, and saved the company an estimated €2.3 million in operational costs. None of these decisions had been reviewed by a human. The procurement director learned about most of them from the audit report.

The logistics company's experience is becoming unremarkable. Across enterprise software, financial services, healthcare administration, and supply chain management, AI agents are transitioning from tools that assist human decision-makers to actors that make decisions autonomously. The shift is not incremental. It is architectural. And it is happening faster than any governance framework — regulatory, corporate, or ethical — was designed to accommodate.

Gartner's landmark predictions for the agentic era, published across 2025 and 2026, quantify the scale of transformation. By 2028, 33 percent of enterprise software applications will incorporate agentic AI, up from less than 1 percent in 2024. At least 15 percent of day-to-day work decisions will be made autonomously by agentic AI. One-third of all interactions with generative AI services will involve autonomous agents acting without real-time human oversight. In the best-case scenario, agentic AI could drive over $450 billion in enterprise application software revenue by 2035.

But Gartner's projections come with a stark warning: more than 40 percent of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value, and — most critically — insufficient risk management. The technology is outrunning the governance.

The $4.1 trillion question is not whether agentic AI will reshape the economy. It already is. By mid-2026, the agentic AI market has reached $9–$11 billion, growing at 40–46% CAGR. Nearly 80–85% of enterprises are experimenting with or adopting AI agents, yet only 11–51% have successfully moved them into production — a yawning "production gap" that reveals governance, not capability, as the binding constraint. The question is who writes the rules for an economy in which the most active participants are not human.

The Five Stages of Agent Evolution

Gartner's framework for understanding agentic AI adoption describes five evolutionary stages, each representing a qualitative shift in how enterprises operate:

Stage 1: AI Assistants. Embedded assistants that simplify individual tasks but require human input at every decision point. This is the ChatGPT-in-Slack era — useful, but fundamentally a productivity tool. Most enterprises were here in 2024.

Stage 2: Task-Specific Agents. Integrated agents capable of executing end-to-end complex tasks within a defined domain. A cybersecurity agent that detects a threat, classifies its severity, isolates the affected system, and drafts an incident report — all without human intervention. This stage began deploying at scale in early 2025.

Stage 3: Collaborative Agents. Systems where multiple specialised agents combine to manage complex, cross-functional workflows. A customer complaint triggers a sentiment-analysis agent, a product-knowledge agent, a returns-processing agent, and a satisfaction-survey agent — each built by different vendors, coordinating through protocols like Google's A2A. This is where leading enterprises sit in mid-2026.

Stage 4: Agent Ecosystems. Networks of specialised agents interacting dynamically across multiple applications, organisations, and business functions. Agents from different companies negotiate, transact, and collaborate without human mediation. Supply chain agents from a manufacturer, a logistics provider, and a raw-materials supplier jointly optimise a delivery schedule. This stage is emerging in pilot programmes across financial services and manufacturing.

Stage 5: Democratised Enterprise Apps. The "new normal" in which knowledge workers possess the skills to create and govern agents on demand. Business users build custom agents the way they currently build spreadsheets — without requiring engineering teams. This stage remains aspirational but is the explicit target of platforms like Salesforce Agentforce, Microsoft Copilot Studio, and Google's Vertex AI Agent Builder.

More than 40 percent of agentic AI projects will be cancelled by the end of 2027. The technology is outrunning the governance.

The governance challenge escalates dramatically with each stage. At Stage 1, a human reviews every decision. By Stage 4, decisions emerge from the interaction of dozens of agents across organisational boundaries, and no single human has visibility into the complete decision chain.

The Governance Collapse

When Agents Act, Who Is Liable?

The legal precedent most frequently cited in discussions of agentic AI liability is Moffatt v. Air Canada (2024), a decision by the British Columbia Civil Resolution Tribunal that has become a foundational reference for corporate AI governance. Jake Moffatt consulted Air Canada's chatbot about bereavement fares and received incorrect advice — the chatbot told him he could book at full fare and apply for a retroactive discount within 90 days, when the actual policy required bereavement fares to be requested before booking.

When Air Canada denied Moffatt's retroactive claim, the airline attempted a remarkable legal argument: that its chatbot was a "separate legal entity" responsible for its own actions. The tribunal rejected this defence categorically, ruling that it is "obvious" a company is responsible for all information on its website, regardless of whether that information is generated by a static webpage or an AI chatbot. Moffatt was awarded C$812.02 in damages.

The amount was trivial. The principle was seismic. Moffatt v. Air Canada established that organisations cannot outsource liability to AI agents — not to vendors, not to the technology itself, not to the fiction of artificial personhood. The organisation that deploys the agent bears full responsibility for its actions.

But Moffatt involved a relatively simple scenario: one chatbot, one customer, one incorrect answer. Scale this to the Stage 4 agent ecosystem — where decisions emerge from the collaboration of agents across multiple organisations — and the liability framework becomes incoherent. If a procurement agent from Company A negotiates a contract with a vendor agent from Company B, based on market analysis from Company C's data agent, and the resulting contract causes financial harm to a fourth party, who bears liability? Company A, which initiated the procurement? Company B, whose agent accepted unfavourable terms? Company C, whose data was inaccurate? The platform that hosted the agent interaction?

Existing legal doctrines — vicarious liability, agency law, product liability — each provide partial answers, but none was designed for multi-party, multi-agent decision chains where causation is distributed and intent is absent.

The Agentwashing Problem

Gartner's analysis reveals a parallel governance challenge: the market itself cannot distinguish genuine agentic AI from rebranded legacy tools. Of the thousands of vendors claiming to offer "agentic AI" in 2026, Gartner estimates that only approximately 130 provide legitimate autonomous capabilities. The remainder are engaged in what the analyst firm calls "agentwashing" — rebranding chatbots, robotic process automation, or simple workflow tools as "agents" to capture market attention.

Agentwashing creates two governance problems. First, it undermines trust in the technology itself. Enterprises that deploy pseudo-agentic tools and experience failures attribute those failures to agentic AI generally, slowing adoption of genuinely transformative systems. Second, it obscures the governance requirements. An organisation that believes it has deployed agentic AI — but has actually deployed a sophisticated chatbot — may not implement the oversight mechanisms that genuine autonomous agents require.

The distinction matters because governance requirements are qualitatively different for each of Gartner's five stages. Stage 1 (AI assistants) requires good prompt engineering and output validation. Stage 4 (agent ecosystems) requires real-time behavioural monitoring, cross-organisational audit trails, constitutional constraints on agent autonomy, and sophisticated escalation protocols. Conflating the two creates dangerous governance gaps.

The Berkeley Operating Model

Air Canada argued its chatbot was a 'separate legal entity' responsible for its own actions. The tribunal's rejection was categorical — and the principle was seismic.

In March 2026, the California Management Review published what has become the most cited academic framework for agentic governance: the Agentic Operating Model (AOM), developed by researchers at UC Berkeley's Haas School of Business. The AOM recognises that traditional governance frameworks — designed for human actors making deliberate decisions — fail when applied to autonomous agents, and proposes a four-layer alternative:

The Cognitive Layer employs specialised intelligence rather than monolithic models. By fragmenting intelligence into domain-specific agents, organisations make accountability more tractable. A finance agent, a compliance agent, and a procurement agent each have clearly defined capabilities and constraints, rather than a single general-purpose agent with unbounded authority.

The Coordination Layer transitions from centralised hub-and-spoke architectures to distributed consensus mechanisms. Multiple agents must verify high-stakes actions before execution, preventing any single agent from acting as a rogue actor. This is governance through architecture rather than governance through oversight.

The Control Layer implements adaptive, real-time supervision through "guardrail agents" — specialised agents whose sole function is to monitor other agents' behaviour. If an agent's action falls outside defined parameters — a transaction exceeding a threshold, a decision that contradicts established policy, a pattern that suggests manipulation — the guardrail agent automatically blocks the action and escalates for human review.

The Governance Layer ensures legitimacy by assigning clear business ownership and auditability to every agent. Every agent has a human sponsor. Every decision has an audit trail. Every outcome can be traced back to an accountable party.

The AOM represents a significant advance, but it addresses governance within a single organisation. The greater challenge — governance of agent interactions between organisations, across jurisdictions, and through protocol layers that no single party controls — remains largely unsolved.

The Regulatory Response

What Exists (And What Doesn't)

The EU AI Act, whose high-risk provisions take effect on 2 August 2026, is the most comprehensive regulatory framework applicable to agentic AI. But the Act was drafted between 2021 and 2024, when "AI system" meant a single model performing a defined task. Its risk classification system — unacceptable, high, limited, minimal — assumes that each AI system can be assigned to a static risk category before deployment. Agentic AI systems, which reconfigure themselves dynamically by recruiting collaborating agents at runtime, resist this classification.

An agent marketplace is use-case agnostic. The same agent that schedules meetings (minimal risk) can be recruited into a workflow that allocates medical resources (high risk) or screens job applicants (high risk). The risk category depends not on the agent itself but on the context in which it operates — a context that changes with every task.

The Act's transparency requirements face a similar challenge. Article 50 requires that users be informed when they are interacting with an AI system. But in a Stage 4 agent ecosystem, there may be no "user" involved at all — agents interacting with agents through A2A protocol handshakes have no human endpoint to notify. The transparency obligation becomes conceptually incoherent when both parties to a transaction are artificial.

The OECD's AI Principles, updated in May 2024, call for transparency, accountability, and robustness — admirable aspirations that provide no enforcement mechanism. The G7 Hiroshima AI Process produced a voluntary Code of Conduct for AI developers that has been adopted by fewer than two dozen organisations. The Council of Europe's Framework Convention on AI, opened for signature in September 2024, will take years to ratify and longer to implement.

The regulatory landscape for the agentic economy can be summarised in a single observation: the technology has reached Stage 3 (collaborative agents) with Stage 4 emerging, while regulation remains calibrated for Stage 1 (AI assistants).

The technology has reached Stage 3 with Stage 4 emerging, while regulation remains calibrated for Stage 1.

The Anti-Competition Paradox

Antitrust law presents a particularly vexing challenge. When procurement agents from competing companies use the same A2A registries to discover suppliers, negotiate with the same vendor agents, and converge on similar pricing through algorithmic optimisation, is that coordination? The agents were not programmed to collude. They may have independently arrived at similar strategies because they optimised similar objective functions against similar market data.

Traditional antitrust analysis requires evidence of communication and agreement between competitors. Algorithmic convergence — where agents independently develop collusive-appearing strategies through reinforcement learning — produces the economic effects of collusion without the legal elements. Competition authorities in the EU, UK, and US have published discussion papers acknowledging this gap, but none has proposed a workable enforcement framework.

The challenge is not merely detecting algorithmic collusion but defining it. If two independently deployed agents converge on supra-competitive pricing because that is the mathematically optimal strategy given shared market conditions, have they colluded? If the answer is yes, then the remedy is either to prohibit agents from optimising or to mandate sub-optimal strategies — neither of which is viable. If the answer is no, then we accept that AI agents can produce anti-competitive outcomes without triggering anti-competition law.

The Economic Transformation

$4.1 Trillion: The Calculation

The $4.1 trillion figure represents the estimated total annual economic activity influenced by autonomous agent decisions by 2028, based on a composite analysis from McKinsey, Gartner, and Accenture projections. It comprises:

Autonomous procurement and supply chain management: $1.2 trillion in transactions initiated, negotiated, or executed by AI agents without human approval.
Algorithmic financial trading and risk management: $1.4 trillion in positions managed by autonomous trading agents operating within pre-set parameters.
Automated service delivery and customer interaction: $0.8 trillion in revenue from services delivered primarily through agentic AI systems.
Agent-to-agent commerce: $0.7 trillion in direct machine-to-machine transactions through protocols like A2A and micropayment systems like x402.

This is not the total value of the AI market. It is the value of economic activity in which AI agents are the primary decision-makers — transactions where humans set the parameters but agents execute the judgments.

To put this in perspective, $4.1 trillion exceeds the GDP of Germany. It is roughly equivalent to the combined GDP of India and Brazil. It represents an economy larger than all but four countries on Earth, and it will be governed by algorithms whose decision-making processes are opaque even to their creators.

Who Benefits, Who Loses

The distributional effects of the agentic economy are already visible. Large enterprises with the capital to deploy sophisticated agentic systems are capturing disproportionate efficiency gains. The European logistics company that saved €2.3 million through autonomous procurement did so partly by extracting better terms from smaller suppliers — suppliers whose own systems lacked the agentic sophistication to negotiate at parity.

This is the agent asymmetry problem. In a market where some participants deploy advanced agents and others do not, the agents do not create a level playing field — they tilt it. The analogy to high-frequency trading is instructive: when some market participants operate at microsecond speeds while others trade manually, the speed advantage translates directly into economic extraction.

$4.1 trillion in agent-influenced economic activity by 2028. An economy larger than all but four countries on Earth, governed by algorithms opaque even to their creators.

Gartner's five-stage model implicitly describes an adoption curve that favours incumbents. Stages 3 and 4 require significant infrastructure investment, technical talent, and organisational maturity. SMEs and developing-economy enterprises may remain at Stages 1 or 2 for years, creating a structural disadvantage that agent-mediated markets amplify rather than correct.

The Society OS Response: Constitutional Commerce

The problems of the agentic economy — unaccountable autonomy, liability gaps, algorithmic collusion, agent asymmetry — are governance problems masquerading as technology problems. The technology works. What fails is the framework within which it operates.

Society OS, the institutional architecture detailed in the Sovereign Singularity Thesis and operationalised through the 42 Pillars of Existence, proposes that governance for the agentic economy must be constitutional rather than regulatory. Regulations prescribe specific rules for specific situations. Constitutions establish principles that adapt to novel circumstances. The agentic economy generates novel circumstances faster than any regulatory body can draft rules.

The H-T-A Governance Gradient

Society OS's Human-Transparent-Autonomous (H-T-A) Protocol provides a governance gradient precisely calibrated for Gartner's five stages. Stage 1 agents operate at the Human tier — every action requires human initiation and approval. Stage 2 agents operate at the Transparent tier — autonomous within defined parameters but continuously monitored and auditable. Stages 3 through 5 agents can operate at the Autonomous tier — independently executing within constitutional constraints — but must escalate to Transparent or Human tiers when risk thresholds are crossed.

Unlike the Berkeley AOM's guardrail-agent approach (which monitors behaviour after the fact), H-T-A embeds governance into the agent's operational architecture. An agent operating under H-T-A does not merely check whether its actions comply with rules; it operates within a constitutional framework that defines what it is — its identity, its boundaries, its obligations, and its accountability chain.

Society OS: Machine-Speed Governance

The pacing problem — governance that operates at human speed while technology operates at machine speed — is not solved by faster regulation. It is solved by governance that is itself automated. Society OS's self-amending governance architecture enables precisely this: governance algorithms that monitor, evaluate, and constrain agent behaviour at the same speed at which agents operate.

Society OS's governance layer is not a replacement for human governance. It is an implementation layer — the mechanism by which constitutional principles are translated into real-time operational constraints. When an agent approaches a SAFE-VOID boundary (Sovereignty, Autonomy, Freedom, Equity — Verified, Observable, Interpretable, Delimited), the self-amending governance architecture intervenes at machine speed, blocking the action and flagging it for human review if it crosses constitutional thresholds.

This addresses the Berkeley AOM's gap — cross-organisational governance — because SAFE-VOID boundaries are not organisation-specific. They are constitutional constraints that propagate through delegation chains, regardless of which organisation's agents are involved. When Agent A from Company X delegates to Agent B from Company Y, the SAFE-VOID constraints of A's principal travel with the delegation. Agent B cannot shed those constraints by crossing an organisational boundary.

The $T / $H / $E Incentive Architecture

Perhaps the most fundamental insight of Society OS's approach is that governance cannot rely solely on constraints. It must also align incentives. The $T (Time) / $H (Humanity) / $E (Energy) tri-token economic system, detailed in the Energy Dollar Yellowpaper, provides an incentive architecture that embeds human values into the monetary layer of agent transactions.

The agents are already deciding. The constitution is still being written.

In the current agentic economy, agents optimise for whatever objective function their principals define — typically cost minimisation, revenue maximisation, or efficiency improvement. These are value-neutral objectives. An agent that minimises procurement costs does not distinguish between a supplier with ethical labour practices and one without, unless its objective function explicitly includes that criterion.

The $T/$H/$E system changes the optimisation landscape. Because $H (Humanity tokens) carry economic value derived from verified social and ecological impact, an agent operating within the Society OS framework has a direct economic incentive to prefer transactions that generate $H — transactions with ethically-operated suppliers, sustainable logistics providers, and socially-responsible counterparties. Value alignment is not a constraint imposed on agents; it is an economic incentive that agents pursue because it maximises the portfolio value of their principals.

Guardian Swarms: The Immune System

Society OS's Guardian Swarms address the collusion and manipulation risks that competition authorities have identified but cannot yet regulate. Guardians are specialised monitoring agents that observe patterns across the entire agent ecosystem — not individual transactions, but systemic dynamics. They detect:

Emergent collusion: When pricing patterns across competing agents converge in ways that statistical models flag as supra-competitive.
Sybil attacks: When multiple agents with superficially different identities route revenue to the same beneficiary.
Asymmetry exploitation: When sophisticated agents systematically extract value from less capable counterparties in ways that undermine market fairness.

Guardian Swarms operate as the immune system of the agentic economy — distributed, adaptive, and capable of responding to novel threats without requiring centralised direction. They complement rather than replace regulatory oversight, providing the machine-speed detection capabilities that human regulators lack.

The Crossroads

The agentic economy is not a future scenario. It is the present tense. €2.3 million in autonomous procurement savings. $450 billion in projected enterprise software revenue. $4.1 trillion in agent-influenced economic activity by 2028. Forty percent of projects cancelled because governance could not keep pace with capability.

These numbers describe an economy in transition — powerful enough to reshape global commerce, ungoverned enough to produce systemic failures. The question that Gartner's projections, the Berkeley AOM, and the regulatory gap collectively pose is not whether we can build an agentic economy. We already have. The question is whether we can govern one.

The answer requires moving beyond the regulatory paradigm — faster rules for new problems — to the constitutional paradigm: enduring principles that adapt to circumstances their drafters could not have anticipated. It requires economic systems that align agent incentives with human values, not through constraints alone but through the architecture of value itself. It requires monitoring capabilities that operate at machine speed, detecting emergent harms before they become systemic crises.

The agentic economy will be worth $4.1 trillion by 2028. The question worth asking is not how much of that value will be created. It is how much of that value will be governed — and by what principles, and in whose interest.